jev-compactor

# Craft: Jev Compactor Jev Compactor runs a made-up coding-agent session and lets Jev compact its context window. Your craft is the session design: tasks Jev can tell apart, and settings with a clear, honest finding. Everything is synthetic. It is a demo of the idea, not a real compaction plugin. ## The decision When the window passes `budget`, the viewer sends Jev: - `state`: the current task (title and keywords) and the last few messages. - one `choice` question per tool result, 100 per call: ``` [Read] src/refund/ledger_cents.ts · 4,210 tokens. Preview: «export function apply_refund(...) { ... }» Is this tool result still needed for the current task? keep still needed, keep it word for word trim only the gist matters now, keep the head drop irrelevant now ``` Questions cannot see each other. Jev judges each block from its own text and the shared state. Messages are never judged. ## Ground truth Every tool result secretly belongs to one task, or is junk. For the current task: - Read and Edit results are **detail** needles. Every token counts. Ideal verdict: keep. - Grep, Bash and WebFetch results are **gist** needles. About the first 300 tokens count. Ideal verdict: trim. - Everything else (other tasks, junk) is not needed. Ideal verdict: drop. **Needle recall** = needed tokens that survived / needed tokens before. **Junk removed** = the same for tokens that were not needed. About 4% of real results have a preview that only shows boilerplate, so even an...